Diagnostic performance of FDA-cleared and CE-marked AI systems for mammography screening: a systematic review and meta-analysis.
Authors
Affiliations (2)
Affiliations (2)
- Faculty of Medicine and Health, The University of Sydney, Camperdown 2050, NSW, Australia; Akademi Teknik Radiodiagnostik dan Radioterapi Bali, Denpasar 80225, Bali, Indonesia. Electronic address: [email protected].
- Faculty of Medicine and Health, The University of Sydney, Camperdown 2050, NSW, Australia.
Abstract
To evaluate the diagnostic performance of FDA/CE-cleared AI systems for mammography screening and assess their roles across clinical workflows. A systematic search was conducted in Scopus, Embase, and PubMed for full-text articles published from January 2018 to July 2024 evaluating commercially available FDA/CE-cleared AI for breast cancer detection using digital mammography (DM) and/or digital breast tomosynthesis (DBT). Included studies specified the AI system and reported diagnostic performance metrics (e.g. sensitivity, specificity, AUC) and/or clinical impact (e.g. cancer detection rate [CDR], recall rate, workload). Study quality was assessed using a modified QUADAS-2 tool. A random-effects meta-analysis pooled standalone AI performance (sensitivity, specificity, and AUC); other workflow roles were synthesised narratively. In total, 42 studies were included (39 retrospective; 3 prospective): 33 evaluated DM, 6 DBT, and 3 both DM and DBT. Risk of bias was frequently high, particularly for flow and timing. In DM, pooled standalone AI AUC was 0.89 (95% CI: 0.85-0.92), with sensitivity of 76.3% (95% CI: 66.2-84.2%) and specificity of 89.6% (95% CI: 84.5-93.2%). Prospective evaluations of AI-integrated workflows, including triage and second-reader replacement, showed non-inferior cancer detection with significant workload reduction compared with standard double reading. A single prospective study of AI as an additional reviewer reported additional cancer detection with minimal increase in recalls. FDA/CE-cleared AI systems demonstrate promising diagnostic performance in breast screening. The strongest evidence supports workflow configurations that reduce radiologist workload, particularly triage and independent/supporting reader models, while prospective evidence supporting complete replacement of radiologists remains lacking.